ieeLetter_2column

jamaicacooperativeAI and Robotics

Oct 17, 2013 (4 years and 22 days ago)

69 views

Support Vector Machine based Inspection of
Solder Joints using a Circular Illumination

T. S. Yun, K. J. Sim and H. J. Kim

An approach to inspection of solder joints using support vector
machines(SVMs) and a tiered circular illumination technique is
propose
d. The illumination technique gives a visual cues to infer
the features of 3D shape for the solder joint surface. The extracted
features are used to classify the solder joint by SVM classifier.
Experimental results show the effectiveness of the proposed
me
thod.


Introduction
: The inspection of the solder joints of surface mounted
devices is crucial in assembling a reliable product. Several solder
joints inspection methods have been previously proposed using
various illuminations and classification technique
s [1][2]. However,
the boundaries between the quality of solder joints are very
ambiguous, therefore, classification
using

these techniques is very
time consuming and requires many training samples. Support vector
machines (SVMs) have been recently propose
d as a method for
pattern classification and nonlinear regression [3]. For several
applications, SVMs have been shown to provide a better
generalization performance than traditional techniques, including
neural networks [4]. This letter proposes an efficie
nt method of
solder joints inspection that uses a SVMs and three
-
tiered circular
illumination approach to infer the characteristics of the 3D shape.
The solder images obtained by the circular illumination are then
used to extract characteristic features. F
inally, each solder region is
classified into one of several pre
-
defined types using an SVM
classifier. Experimental results suggest that the proposed method
could be applied to classifying various shapes of specular solder
joints in real situations


The C
ircular Illumination and Feature Extraction
: As shown in Fig.
1, the experimental apparatus consisted of a CCD color camera and
three color circular lamps(blue, red and green) with different
illumination angles. Since solder surface has a specular nature,
the
surface patches of solder joints having the same slope show the same
intensity patterns called iso
-
inclination contours in a camera image.
The highlight pattern gives us good visual cues for inferring 3D
shapes of the specular surface. Fig. 2 illustrat
es examples of the
solder image obtained by the camera. The solder joints being
inspected were divided into four classes according to their quality:
Good, Excess, Insufficient, and No solder.



Fig. 1

The experimental apparatus








a b c d

Fig.2

The types of solder joint

a

excess

b

good

c

insufficient

d

no solder


The images obtained by the CCD camera are then preprocessed to
extract the characteristic fea
tures including the average intensity
value and percentage of highlights. The characteristic features can be
defined as follows:



(1)






where

is a six
-
dimensional feature vector, subscript
i

and
k
indicate the feature vector elements 1,2,3 and 4,5,6,
respectively,

is the intensity of the image in the
l

color
frame
(1=red, 2=green, 3=blue ),

is the thresholded value of the
image in each color frame, and
N

is the number of pixels in the solder
region . The value of threshold is determined empirically to be 170.


Classification using Suppor
t Vector Machines
: Given a set of points

that belong to one of two classes, an SVM finds the
hyperplane by leaving the largest possible fraction of points of the
same class on the same side, while maximizing the distance of either
c
lass from the hyperplane. The hyperplane
-
called the Optimal
Separating Hyperplane(OSH)
-

can be represented as:



(2)

where

is the weight vector, the ve
ctor z represents the
image
induced in the feature sp
ac
e due to
the kernel
,
b

is the bias

and nonnegative variables
are
Lagrange multipliers

can be
obtained by standard quadratic optimization technique[3]
.

A tr
ained
SVM determines the class of a test pattern according to the sign of
eqn. 2. For linearly non
-
separable problems, such as the classification
of a solder joint , the OSH can be identified through nonlinear kernel
mapping into high dimensional space. Th
e kernel used in our
experiment is polynomial of degree 3 defined by:



(3)


Since methods using an SVM usually classify only two pattern
classes, this must be extended

to apply to the classification of solder
joints. The approach taken to formulate this multi
-
class classification
problem was to consider it as a series of binary classification
problems. Accordingly, four classifiers were constructed, with one
for each cl
ass ( e.g. the good classifier separates patterns belonging to
the good class from all the others). For a particular test pattern, the
classifier with the highest output value was selected as the winner and
the corresponding class label then assigned. Fig.
3 shows the
architecture of solder joint classifier. The input to the classifier is
six
-
dimensional feature vector extracted from the solder joints image.
Each element of feature vetor was normalized to [0,1]. The number of
kernel
n
r

for each
r
th

SVM

is th
e same as the number of support
vectors obtained when training for classifying class
r

and the other
classes.


Fig. 3

The Architecture of the solder joints classifier
.


Experimental Results
: In order to verify the effectiveness

and
robustness of the proposed classifier, experiments were performed on
various solder joints. First, the performance of a classifier using an
SVM was evaluated, and then the effect of the size of the training
data and its performance were addressed. The

experiments were
carried out on an IBM
-
PC pentium with Windows

98 using VC++.
The capture board used was Dipix PicPort, and the inspection field
of view(FOV) is set 10mm x 10mm on a PCB, within which five or
six solder joints are usually imaged. Each sam
ple image is a
windows subimage fitted to a solder joint with a size of 50 x 100
pixles. Sample data on 402 solder joints was collected from
commercially manufactured PCBs. 201 solder joint samples were
used for training and 201 samples were used for testi
ng. Table 1
presents the classification results for the testing data. For class
Insufficient and No Solder, all test data are correctly classified. For
class Excess, only two samples are misclassified into the class Good,
resulting in a success rate of 96
.07%. For class Good, only one
sample is misclassified into the class Excess, resulting in a success
rate of 98%. To elucidate the merits of the proposed method,
comparative studies are made with k
-
meams and Back
propagation(BP) algorithm, and the results

are shown in Table2. The
k
-
means classifier is easy to implement and very fast, but has row
classification rates.It takes BP classifier long time to train the large
samples and can

t classify between the class good and excess well.


Table1:

Calssification

results of testing data

Types

Excess

Good

Insuffic
ient

No
Solder

No.of
samples

Excess

49

2

0

0

51

Good

1

49

0

0

50

Insufficient

0

0

50

0

50

No Solder

0

0

0

50

50


Table 2:

Comparison results with other classifier

Types

k
-
means(k=4)

BP

SVMs

Excess

88
.23

92.16

96.07

Good

92

96

98

Insufficient

94

100

100

No Solder

100

100

100


In order to evaluate the performance of an SVM that maximizes the
distance of two classes from the hyperplane, an SVM was trained for
several sets of data with different sizes
. As shown in table 3, the
SVM produced a good classification result even though the sample
size was so small.


Table 3:

Classification performance for different sizes of training data

Number of training data

50

100

150

200

Classification rate

96

98

98

9
9


Conclusions
: This letter has presented an efficient method of solder
joints inspection that uses a SVMs and three
-
tiered circular
illumination approach to infer the characteristics of the 3D shape.
Experimental results show that the proposed method pro
duces a high
classification rate in the linearly non
-
separable problem of
classifying solder joints. Additionally, this method can produce a
good result even with a small size of training data due to the nature
of maximizing the distance of either class fr
om the hyperplane.
Further work will involve the utilization of the tilt angles of the
solder surface in the proposed method, plus an investigation of the
optimal position and number of lights that can estimate the
reflectance parameters and diffuser ratio

of a specular component.


References


1.

S.K.Nayar, A.C.Sanderson, L.E.Weiss, D.D.Simon,

Specular surface
inspection using structured highlight and gaussian images,


IEEE Trans.
Robotics and Automation
, Vol. 6, No. 2, pp. 108
-
218, 1990.

2.

T.H.Kim, T.H.Cho,

Y.S.Moon, S.H.Park,

Visual inspection system for
the classification of solder joints,


Pattern Recognition

, Vol. 32,
pp.565
-
575, 1999.

3.

HAYKIN, S.: Neural network: a comprehensive foundation, second
edition, Prentice Hall, 1999.

4.

SCHOLKOPF, B., SUNG, K.,
BURGES, C., GIROSI, F., NIYOGI, P.,
POGIO, T., and VAPNIK, V.:

Comparing support vector machines with
Gaussian kernels to radial basis function classifiers

, A. I. Memo 1599.